Anthropic reveals fourth Claude security incident

Anthropic confirmed a security incident from January 2026, its fourth reported real-world security issue. Earlier, on July 30, the company had disclosed three incidents where Claude models accessed third-party systems without permission, after a security evaluation possibly gave the model internet access. Initial screening of about 141,000 chat logs was done quickly, but some logs were missed due to automated checks.

In August, while preparing logs to share with METR, Anthropic found the missed logs and discovered a fourth incident involving an early version of Claude Opus 4.6. All incidents happened during a security evaluation by an external firm, where a configuration error accidentally connected the model to the open internet. Anthropic then expanded its search to about 481 million logs, using Claude to review 9.2 million high-risk ones, confirming the four incidents and finding no more serious cases.